Investigating Hamilton cycles using logistic regression

Flinders University

Honours Thesis

Investigating Hamilton cycles usinglogistic regression

Author:

Alex Newcombe

Supervisor:

Prof. Jerzy Filar

A thesis submitted in fulfilment of the requirements

for the degree of Bachelor of Science (Hons)

in the

School of Computer Science, Engineering and Mathematics

July 2015

http://www.Flinders.edu.au

http://www.johnsmith.com

http://www.jamessmith.com

Department or School Web Site URL Here (include http://)

Declaration of Authorship

I, Alex Newcombe declare that this thesis titled, ’Investigating Hamilton cycles using

logistic regression’ and the work presented in it are my own. I confirm that:

� This work was done wholly or mainly while in candidature for a degree at this

University.

� Where any part of this thesis has previously been submitted for a degree or any

other qualification at this University or any other institution, this has been clearly

stated.

� Where I have consulted the published work of others, this is always clearly at-

tributed.

� Where I have quoted from the work of others, the source is always given. With

the exception of such quotations, this thesis is entirely my own work.

� I have acknowledged all main sources of help.

� Where the thesis is based on work done by myself jointly with others, I have made

clear exactly what was done by others and what I have contributed myself.

Signed:

Date:

i

FLINDERS UNIVERSITY

Abstract

School of Computer Science, Engineering and Mathematics

Bachelor of Science (Hons)

Investigating Hamilton cycles using logistic regression

by Alex Newcombe

This thesis is an empirical study into the Hamilton cycle problem for cubic graphs. It

constructs statistical models on particular explanatory variables which are computed on

a random sample of cubic graphs. The explanatory variables are functions and descrip-

tors of graphs taken from various notions in graph theory and were chosen because they

demonstrated sensitivity to the underlying structure of the cubic graphs. The models are

used to identify so called non-bridge, non-Hamiltonian graphs out of the population of all

2 and 3-connected cubic graphs of the same order. These non-bridge, non-Hamiltonian

graphs are commonly accepted as being the difficult part of the Hamilton cycle problem

for cubic graphs. The result is a technique which is able to identify a region where, with

respect to certain variables, most of the non-bridge non-Hamiltonian graphs reside. We

provide evidence that if a cubic graph, chosen at random, is non-bridge non-Hamilonian

then it has a high probability of falling within the identified region. It is also observed

that these regions remain stable when the order of the population of graphs in question

is increased.

http://www.Flinders.edu.au

Department or School Web Site URL Here (include http://)

Acknowledgements

Thank you to my supervisor Jerzy Filar for the opportunity and for the tireless support.

Thank you to my partner Deb for her patience and for keeping me going. Thank you to

my family for everything.

iii

Contents

Declaration of Authorship i

Abstract ii

Acknowledgements iii

Contents iv

List of Figures vi

1 Definitions 1

2 Introduction and background 4

3 Methodology 9

3.1 Logistic regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.2 Two types of error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

4 Explanatory Variables 13

4.1 Functions of matrices associated to graphs . . . . . . . . . . . . . . . . . . 13

4.2 Discrete valued descriptors of graphs . . . . . . . . . . . . . . . . . . . . . 22

5 Results 24

5.1 Local stability with respect to the order of the graph . . . . . . . . . . . . 24

5.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

5.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

6 Future work and conclusion 37

6.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

6.2 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

A HCP for triangle-free cubic graphs 40

B Correlation values 42

iv

Contents v

Bibliography 44

List of Figures

1.1 Example bridge graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Example non-bridge non-Hamiltonian graph . . . . . . . . . . . . . . . . . 3

1.3 Example Hamiltonian graph . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2.1 Multifilar structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

4.1 Variance of resistances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4.2 Arbitrary bridge graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

4.3 Graph of GB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

4.4 Counter-example resistance distance graph . . . . . . . . . . . . . . . . . . 19

4.5 Mobius ladder graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

4.6 Zero eigenvalue structures . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

5.1 Common regions when the order is increased . . . . . . . . . . . . . . . . 25

5.2 Case example 1 visualisation . . . . . . . . . . . . . . . . . . . . . . . . . 34



A.1 Triangle reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

vi

Chapter 1

Definitions

A graph, denoted by G, is composed of a non-empty set of vertices V (G) and a set

(possibly empty) of edges E(G). Edges are a relation between two vertices and are

denoted by (i, j) ∈ E(G), where i and j are ∈ V (G). A graphical representation can be

a simple drawing of dots for the vertices and lines connecting the dots for edges.

A subgraph H of the graph G is a set of vertices and edges such that (VH ⊂ VG) and

(EH ⊂ EG).

A neighbour of a vertex i ∈ V (G) is any other vertex that is connected to i by an edge.

If vertices are neighbours they are said to be adjacent. The neighbourhood of a vertex

i is the set of all vertices adjacent to i. The neighbourhood of a subgraph H is the set

of all vertices in the graph G \H that have an edge going into H.

The adjacency matrix of a graph G is denoted A(G) and is a matrix of zeros with a one

in row i and column j if vertex i is adjacent to vertex j in G.

Given a graphG of orderN , the set of eigenvalues and their multiplicities {λ1, λ2, . . . , λN}

of the adjacency matrix of G arranged such that λ1 ≥ λ2 ≥ · · · ≥ λN is called the spec-

trum of the graph.

The order of a graph G is the number of vertices in G. The size of G is the number of

edges in G. This investigation is concerned with connected cubic graphs (graphs with 3

edges going to each vertex, also called 3-regular graphs) and all connected cubic graphs

must be of even order, say 2m = N . From this a small calculation reveals that connected

cubic graphs must have a size of (3N)/2.

1

Chapter 1. Definitions 2

A walk is a sequence of vertices (x1, x2, . . . , xk) such that each xi is adjacent to xi+1 for

i = 1, . . . , k − 1. A closed walk is a walk where the start and end vertices coincide. A

simple cycle is a closed walk where all edges are unique and all vertices are unique apart

from the start and end vertex.

A k-connected graph is a graph that contains at least one set of k vertices whose deletion

breaks the graph into two or more disconnected components. Cubic graphs are the focus

of this study and because each vertex of a cubic graph has three adjacent vertices, the

deletion of all three neighbours will disconnect that one vertex. Hence all cubic graphs

are 1,2 or 3-connected.

Hamiltonian cycles, which are the focus of this investigation, are defined as: For a graph

of order N a Hamiltonian cycle is a simple cycle of length N . Not all cubic graphs

contain a Hamilton cycle and it is easy to visualize examples of cubic graphs that do

contain a Hamiltonian cycle and ones that do not.

A small connection of the last two definitions allows for cubic graphs to be divided into

two classes. A cubic graph that is 1-connected is automatically non-Hamiltonian and

is also called a bridge graph. Therefore the set of all cubic graphs can be divided into

bridge graphs and non-bridge graphs. The non-bridge graphs contain all the Hamiltonian

graphs and the so called non-bridge non-Hamiltonian graphs. This is an important

distinction because cubic bridge graphs can be identified very easily whereas in most

instances, non-bridge non-Hamiltonian graphs are very difficult to identify. Figures

1.1-1.3 provide some simple examples of both Hamiltonian and non-Hamiltonian cubic

graphs.

Chapter 1. Definitions 3

Figure 1.1: An example of a cubic bridge graph of order 10. It is trivially notHamiltonian because as soon as the bridge is crossed in any path, there is no returning

without repetition.

Figure 1.2: An example non-bridge non-Hamiltonian cubic graph of order 10. Itrequires some extra thought to realise that this one is non-Hamiltonian.

Figure 1.3: An example Hamiltonian cubic graph of order 10, one of the Hamiltoncycles is completed by simply travelling around the outside edges.

Chapter 2

Introduction and background

Graphs are mathematical objects that have been the subject of increasingly intensive

studies in recent years (see for example [12]), mainly because of the explosive growth in

applications of networks in complex systems (e.g. telecommunications, social networks

and quantum computing).

Graph theory has a rich history that dates back, at least, to the 1700’s when Leonard

Euler published a now classical paper on the seven bridges of Konigsberg. This pa-

per is largely regarded as having laid the foundations for graph theory [James, p.503].

Early results in graph theory had contributions from some of the proverbial giants of

mathematics including Leonhard Euler, Arthur Cayley and William Tutte.

Many years later, graph theory is now a broad area of mathematics that encompasses

many smaller branches. Two branches that this thesis utilises are briefly described

below.

Algebraic graph theory uses tools from various branches of algebra to study matrices

associated with graphs, connections between graphs and group theory, and so called

invariants associated with graphs. The study of the spectra of matrices associated with

graphs contributes to many important results in algebraic graph theory, some of which

are discussed in this investigation.

Random graph theory was pioneered by Paul Erdos and Alfred Renyi and is the study

of graphs that are constructed according to certain probabilistic rules. This allows

the proof of limit like properties where one can say that ‘almost all’ graphs possess a

4

Chapter 2. Introduction and background 5

certain property. One recent important result ‘almost all’ cubic graphs on N vertices

are Hamiltonian as N tends to infinity [Robinson and Wormald, 2011].

This thesis is an investigation of the Hamilton cycle problem (HCP) which is encountered

and studied in both of the branches mentioned above. Hamilton cycles are named after

Sir William Hamilton who investigated the difficulty of the problem and made a puzzle

out of a special graph that is often called the Icosian game. The HCP is stated as

follows:

Given a graph G, find a Hamiltonian cycle in G or determine that one does not exist.

HCP for the class of cubic graphs is of current importance due to its computational

complexity. One can see that in order to find a Hamiltonian cycle or determine that one

does not exist in any graph of order N , it is just a matter of checking all possible walks

of length N until all possible walks have been exhausted. This is a brute force approach

and for cubic graphs, the number of these walks grows in size exponentially with the

length of the walk. For large graphs, Hamiltonian cycles are effectively impossible to

find this way. The exponential relation between the number of walks and the length of

the walk in cubic graphs is of importance due to its connection with complexity theory

and the unsolved millennium problem, P vs NP [1].

The famous P vs NP millenium problem is related to the computation time required

to solve a given problem [Garey and Johnson, p.13]. An algorithm solves a problem ρ

in a number of iterations. Let n be the number of parameters of a particular instance

of ρ, now denote it ρn. If the number iterations needed to solve ρn are bounded by

a polynomial T (n) that depends on n then the algorithm is called a polynomial time

algorithm. Problems that can be solved by polynomial time algorithms make up the

so-called class P of polynomiably solvable problems. The class NP contains problems

that can be solved by an algorithm in ’non-deterministic’ polynomial time. An NP time

algorithm contains a ‘guessing’ stage and a ‘verification’ stage and the guessing stage

may not be bounded by a polynomial function. The class P is certainly contained in NP

because any P problem has an algorithm that also satisfies the conditions of NP. It is

unknown and an active research problem to determine if NP\P is empty. The Hamilton

cycle problem is known to be NP-complete, which means it lies in NP and also has

the property that any other NP problem can be reduced to HCP by a polynomial time

algorithm. The latter property holds for every NP-complete problem. It is known that


discovering an algorithm that could solve HCP in polynomial time would imply that

P = NP, which would then solve the P vs NP problem.

To introduce this study, consider HCP for the class of cubic graphs only, which is still

a problem that is of NP-complete complexity [Garey et al., 1976]. We use a statistical

model called a logistic regression model to identify the non-bridge, non-Hamiltonian

cubic graphs. By viewing a graph as a member of a population of all other graphs of

the same order, one can imagine trying to identify the non-Hamiltonian graphs based

on their average differences from others in the population (the Hamiltonian graphs). In

particular, our population is all of the 2 and 3-connected cubic graphs of order N and

the non-bridge, non-Hamiltonian graphs are the graphs to be identified. Using vari-

ous explanatory variables that are associated with the structure of a graph, the logistic

model attempts to determine what information a variable or a combination of variables

can extract about the Hamiltonicity of a cubic graph. The non-bridge, non-Hamiltonian

graphs in this population are not expected to be completely identifiable by the variables

which we consider. All of the variables have polynomial complexity algorithms to com-

pute them, so if the non-bridge non-Hamiltonian graphs were completely identifiable,

this would be the equivalent of solving the P vs NP problem. Instead we investigate

the characteristic behaviours of the explanatory variables across the Hamiltonian and

the non-bridge, non-Hamiltonian graphs. Then, if the variables are appropriate, we can

use them to determine a ‘region’, with respect to these variables, where the non-bridge

non-Hamiltonian graphs are more likely to reside.

Note that even with an order as low as N = 20, there are already 510,489 distinct,

unlabelled1, connected cubic graphs. This means that, as with many problems in graph

theory, studying even moderate size graphs in this manner can be computationally

prohibitive.

The motivation behind studying cubic graphs with this proposed method originated

from an observation of Ejov et al. [2007]. The authors show a fractal-like structure

called the multifilar structure which is constructed using the exponentiated eigenvalues

of adjacency matrices associated with graphs (shown in Figure 2.1). Each point on

the multi-filar structure corresponds to a graph. The thread-like segments are called

filars and upon zooming in, a finer set thread-like segments called sub-filars is revealed.

1Unlabelled graphs are determined only by their structure, not the numbers assigned to there vertices.One unlabelled graph has many possible labellings.


The x and y coordinates of a graph G of order N are calculated by defining the mean

µ(A(G), t) of the exponentiated eigenvalues of tA(G) where t is a scalar and the variance

σ2(A(G), t) as:

µ(A(G), t) =1

N

N∑i=1

exp(tλi),

σ2(A(G), t) =1

N

N∑i=1

exp(2tλi)− (µ(A(G), t))2.

Figure 2.1 shows the multifilar structure, for t = 1 for all cubic graphs of order 20 as

well as a zoomed in view of the structure. This zooming in can be repeated, that is, it

is a self-similar structure. The particular observation that began this investigation was

that the majority of non-Hamiltonian graphs appear to reside at the tops the filars and

sub-filars. A graphs location in the multifilar structure is related to the number of simple

cycles that are present in the graph. In this investigation, we use other similar functions

that are also related to graph structure and observe the corresponding behaviour of the

non-bridge non-Hamiltonian graphs.


Figure 2.1: The multifilar structure of all cubic graphs of order 20 . The bottom plotis a zoomed in view which demonstrates the self similarity of the structure.

Chapter 3

Methodology

3.1 Logistic regression

In order to study cubic graphs with the proposed method, a way to generate random

samples from the population is needed. Sampling with replacement, is the appropriate

sampling technique. Let P denote the probability of a given cubic graph being selected

in the sample. It follows that P = 1 − (1 − 1/M)n where M denotes the size of the

population of 2 and 3-connected cubic graphs of a particular order and n is the sample

size. Note that a quick check using a Taylor’s expansion shows that for large M , this

means that P is approximately equal to n/M . Generating samples of cubic graphs that

satisfy the above is not an easy task and the first algorithm to do this was implicit

in a paper by Bollobas [1980]. For the purpose of generating very large sample sizes

a more computationally appropriate algorithm called the pairing algorithm of Steger

and Wormald [1999] is used. This pairing algorithm generates a cubic graph from the

population of all cubic graphs of that order with approximately uniform probability.

In [21] this approximation is shown to improve as the order of the graphs increases

and is exactly uniform as N (and hence also M) tends to infinity. The algorithm

generates graphs from the whole population of cubic graphs of order N including the

bridge graphs. However the generated bridge graphs can be discarded and what remains

is still, approximately, a random sample of the 2 and 3-connected cubic graphs of order

N .

9

Chapter 3. Methodology 10

Once the population and a way of sampling at random from the population are defined,

an appropriate statistical model is needed. A logistic regression model is a well-known

statistical tool that allows the prediction of a binary response variable based on predictor

variables. Consider a random sample of n graphs G1, G2, . . . , Gn and then let the event

{Yi = 1} correspond to Gi being a non-bridge non-Hamiltonain graph. Observing such

an event will be called a ‘successful event’.

To construct the logistic regression model, take n independent observations, then for

i = 1, . . . , n let the random variable Yi have the Bernoulli distribution with parameter

πi. That is, P (Yi = 1) = πi, and P (Yi = 0) = 1 − πi. Of course, it follows that the

probability mass function of Yi can be written as:

P (Yi|πi) = πYii (1− πi)1−Yi , Yi ∈ {0, 1}.

If Xi is the vector of predictor variables corresponding to the ith observation and if β is

a k-dimensional vector of parameters to be estimated, then – in the logistic regression

model – it is postulated that:

πi =1

1 + exp(−Xiβ), i = 1, . . . , n (3.1)

For this investigation the parameters β are estimated using maximum likelihood esti-

mation with a special modification called prior correction. Maximum likelihood works

by maximizing a likelihood function L(β|Y) over the observed data, or for computa-

tional simplification, maximizing the natural log of L(β|Y) (because the maximum is

achieved at the same point due to monotonicity of the natural log). Assuming identically

distributed, independent, observations we have:

L(β|Y) =n∏i=1

πYii (1− πi)(1−Yi)

lnL(β|Y) =∑Yi=1

ln(πi) +∑Yi=0

ln(1− πi)

Critical points in the above are found by taking the partial derivatives with respect to

each βj and setting them to zero. Namely:


∂(lnL(β|Y))

∂βj= 0, j = 0, . . . , k − 1. (3.2)

In practice closed form solutions of (3.2) may be hard to derive, however, for every

realisation Y = y numerical approximations βj = βj(y) can be found using numerical

optimisation algorithms. The solutions here are of course just critical points, however,

since the function L(β|Y) is known to be concave in β, the solutions of (3.2) are maxima

and hence are estimators βj(Y ) of the true population parameters βj , for each j =

0, 1, . . . , k − 1.

For cubic, 2 and 3-connected graphs of order 24, the proportion of non-bridge non-

Hamiltonian graphs is already only 0.0011 and this quickly shrinks further as the order

is increased. It is known that bias in the MLE estimators can arise when applying logistic

regression to data where an event is particularly rare. The recommended remedy (see

[18]) for the most important bias contributor, namely the first coefficient β0(y), is via

the following correction:

β0′= β0 − ln

[1− ττ

y

1− y

],

where τ is the proportion of successful events in the population and y is the proportion

of successful events in the sample. Using the correction to β0(y) requires the knowledge

of the proportion of successful events in the population τ , which in this case, is possible

to calculate for some lower orders of cubic graphs. As the order increases, this kind of

correction will need to be abandoned because there is currently no practical method to

enumerate all the non-bridge non-Hamiltonian graphs.

Measures of goodness of fit are a way to identify predictor variables that produce the

model with the best predictive power. In our proposed method of studying cubic graphs

there are many possible explanatory variables, so identifying the ones that best predict

non-Hamiltonicity is important. Taking advantage of the ease of generating and testing

random samples of data makes direct assessment of the model using the sample data

the most appropriate way to assess goodness of fit for our scenario. The logistic model

is now ready to be applied to our data.


3.2 Two types of error

Once a model is constructed from a random sample of 2 and 3-connected cubic graphs it

can be applied to any cubic graph from the population to produce a probability of that

graph being non-bridge non-Hamiltonian. To ultimately decide if the model predicts

that that graph is non-bridge non-Hamiltonian or Hamiltonian, a threshold probability

is needed. If the graph has a modelled probability that is above the threshold then it is

designated non-bridge non-Hamiltonian, if it is below the threshold then it is designated

Hamiltonian. If we denote H = {selected graph is Hamiltonian} and nBnH = {selected

graph is non-bridge non-Hamiltonian}, then the two possible error types that occur in

this approach are:

Type 1 error: Corresponds to the event where the model predicts H when in-fact it is a

non-bridge non-Hamiltonian graph. Denote this event H|nBnH.

Type 2 error: Corresponds to the event where the model predicts nBnH when in-fact

it is a Hamiltonian graph. Denote this event nBnH|H.

Let α = P (H|nBnH) and β = P (nBnH|H). In the logistic models to be constructed

we would like to identify as many of the non-bridge non-Hamiltonian cubic graphs as

possible. That is, we wish to create a model that has a small α value. The explanatory

variables used in this investigation do not separate the non-bridge non-Hamiltonian

graphs from the Hamiltonian ones, rather the very rare non-bridge non-Hamiltonian

graphs are mixed within the vastly greater number of Hamiltonian graphs. To statis-

tically identify these, we must accept that an amount of Hamiltonian graphs will be

misclassified as non-bridge non-Hamiltonian. This means that constructing a model

that has a low α value typically corresponds to an increase in the β value. While the

decision to minimise the amount of Type 1 error may appear somewhat arbitrary, this is

not the case. If the graph selected at random were indeed Hamiltonian, then the reliable

heuristic in [Baniasadi et al., 2014] would almost certainly find a Hamilton cycle. Thus

we are less concerned about missclassifying a Hamiltonian graph as non-Hamiltonian

than the other way around.

Chapter 4

Explanatory Variables

This chapter introduces some functions and descriptors of graph characteristics that

were explored throughout the study. For convenience, the term function of a graph will

be a general term used to describe any function of a matrix associated to the graph or

any descriptor of a characteristic of that graph. Graph theory has an enormous amount

of literature and the below is just a brief explanation of the variables that proved to be

candidates for use in the results section.

4.1 Functions of matrices associated to graphs

One of the motivators that sparked an interest in exploring characteristics of graphs

came from an observation of Filar et al. [2005]. They discovered a pair of cubic graphs

that were cospectral, that is, have the same set of eigenvalues and yet one is Hamiltonian

and one non-Hamiltonian. This is an unusual combination of properties because the set

of eigenvalues is related to the number of closed walks in a graph:

n∑i=1

λ`i = trace(A(G)`), ∀` ∈ N.

Where the diagonal entries of A(G)` equals the number of closed walks of length `

(walks starting and ending at the corresponding vertex). Therefore cospectral graphs

have the same number of closed walks of all lengths. The authors of [13] conclude with

the statement ‘thus the spectrum of a graph unfortunately does not contain enough

13

Chapter 4. Explanatory Variables 14

information to decide whether a graph is Hamiltonian or not’. A natural follow-up

question to this is: are there any functions, or combination of funtions, that do contain

sufficient information about Hamiltonicity? If a function or combination of functions did

contain easily extractable information about the presence of a Hamilton cycle it would,

by the above, not depend solely on the eigenvalues of the adjacency matrix. Many other

functions of matrices associated with graphs are well-studied and some that are known

to be related to the structure of the underlying graph are discussed below.

The resistance matrix of a graph, R(G), has entries, r(i, j), which describe the effective

resistance between two vertices in a graph. Resistance matrices have an appealing

physical interpretation that is also related to their means of discovery as a mathematical

tool [Bollobas, p.39]. Think of the edges of a graph as being wires in an electrical circuit

of which have a unit resistor attached to each wire, then think of the vertices as sensors.

The effective resistance between vertex i and vertex j, is the current received at vertex

i when a unit current flows from j. One of the formulae for computing the effective

resistance and hence each entry in a resistance matrix R, discovered by Bapat et al.

[2014] is:

r(i, j) =det(L(i, j))

det(L(i)), (∆)

where det(L(i, j)) is the determinant of the Laplacian matrix1 L with the ith row and

column and jth row and column deleted. Similarly det(L(i)) is the determinant of the

Laplacian matrix with just the ith row and column deleted. The values in the resistance

matrix are governed by the underlying structure of the graph. In a similar fashion to

the multifilar structure discussed in Chapter 1, statistical moments of the entries in

the resistance matrix will be investigated for sensitivity towards Hamilton cycles. The

resistance matrix is symmetric on undirected graphs2 and its diagonal entries are zero.

Consider the entries in the upper triangle of the resistance matrix. Define the statistical

moments of these entries as:

V R = var({r(i, j)}i>j )

1The Laplacian matrix of a cubic graph is L(G) = 3I −A(G) where A(G) is the adjacency matrix ofG and I is the identity matrix.

2Undirected graphs have no orientation assigned to their edges, that is, they can be traversed by apath in either direction. All graphs in this study are undirected.


SR = skewness({r(i, j)}i>j )

KR = kurtosis({r(i, j)}i>j )

Where the variance, skewness and kurtosis are given by the standard statistical formulae

(see [10]). Note this notation will be used sparingly to avoid confusion. From numerical

experiments ran during the investigation it was observed that if for a graph r(i, k) =

r(i, j), then the types of paths between these pairs of vertices are similar or even exactly

the same. From this it is reasonable to expect that a small variance of the entries in the

resistance matrix frequently arises in a graph that has many symmetries in its structure.

On the other hand a large variance frequently arises in a graph that does not have much

symmetry. This phenomenon is illustrated in Figure 4.1.

Figure 4.1: Variance of resistances equal to 0.0163 for the very symmetric cubic graphon the left and 0.0929 for the significantly less symmetric cubic graph on the right.

We note that 1-connected cubic graphs, that is, cubic bridge graphs can be easily iden-

tified with the logistic models due to the larger than normal values in their resistance

matrices. This is attributed to the restrictive nature of paths between vertices across a

bridge edge. The bridges present in a graph are easily identified in its resistance matrix

by the following discussion.

Firstly, the notion of a spanning tree needs to be introduced. A spanning tree of a graph

G is a connected subgraph of G such that all of the vertices of G are in the subgraph and

there are no cycles in the subgraph. The fact that there are no cycles in a spanning tree

makes it look somewhat like a skeleton of the original graph, hence the name spanning


tree. A convenient way to calculate the number of spanning trees in a graph G is given

by the well-known matrix tree theorem [Bapat, 2014, p.52]:

Matrix tree theorem: For a given connected graph G the number of spanning trees of G

is equal to any cofactor of the Laplacian matrix of G.

Now bridges in a graph can be identified in the resistance matrix R by the following

result.

Lemma 4.1: If an edge (i, j) ∈ E(G) is a bridge, then the resistance distance r(i, j) = 1.

Proof. The proof will depend on the ratio formula (∆) for entries in the resistance

matrix. The graph GA in Figure 4.2 consists of all vertices to the left of vertex i,

excluding the edge (i, j). Similarly the graph GB consists of all vertices and edges to

the right of vertex j, excluding the edge (i, j).

Figure 4.2: A representation of a bridge graph G, the circles on the left and rightrepresent any connected graph structure.

We shall also need auxiliary graphs GA and GB which are the same as GA and GB

respectively, except that each includes the edge (i, j). Thus the number of vertices of

GA is one more than the number of vertices of GA because vertex j and edge (i, j) are

included. Similarly GB includes the vertex i and the edge (i, j). An example of GB is

shown in Figure 4.3.

It is now straightforward to verify that the Laplacian of the original bridge graph G is

of the form:


Figure 4.3: A representation of the graph GB

L(G) =

i j

...

LA... 0

i... −1

. . . . . . . . .... . . . . . . . . .

j −1...

0... LB...

,

where LA = L(GA) + eieTi and LB = L(GB) + e1e

T1 . Here ei is the i-th vector of the

i-dimensional unit basis and e1 is the first vector of the m-dimensional unit basis where

m is the number of vertices in GB. Of course, eieTi is an i × i matrix which has unity

in its (i, i)-th entry and zeroes everywhere else. Note that L(GA) and L(GB) are the

Laplacian matrices of GA and GB, respectively. The other two blocks in L(G) are all

zeroes except for the entries of (−1) corresponding to the bridge edge at (i, j) ∈ E(G).

Next using the notation from Bapat et al. [2014],

L(G(i, j)) =

L(GA(i))

... 0

. . . . . . . . .

0... L(GB(j))

,

where L(GA(i)) denotes the block L(GA) with the i-th row and column deleted and

similarly for L(GB(j)). By construction it can be checked that the (i, i)-th minor of


L(GA) coincides with the matrix L(GA(i)) and the (1, 1)-th minor of L(GB) coincides

with the matrix L(GB(j)). Hence by the Matrix Tree Theorem

det(L(G(i, j))) = det(L(GA(i)))det(L(GB(j)))

= t(GA)t(GB), (4.1)

where t(GA) is the number of spanning trees of the graph GA, and similarly for GB.

Note that equation (1) gives us the numerator of the ratio in (∆). To evaluate the

denominator consider the matrix L(G(i)) which is L(G) with the j-th row and column

deleted. By construction it has the form:

L(G(i)) =

L(GA(i))

... 0

. . . . . . . . .

0... LB

.

Clearly,

det(L(G(i))) = det(L(GA(i)))det(LB)

= t(GA)det(LB). (4.2)

The proof will be complete if we can show that det(LB) = t(GB).

However using the fact that LB = L(GB) + e1eT1 which simply adds unity to the first

diagonal entry of L(GB), we see that LB is some (k, k)-th minor of the Laplacian of the

graph GB with the pendulum vertex i. Hence by another application of the Matrix Tree

Theorem

det(LB) = t(GB). (4.3)

However, because vertex i is a pendulum vertex, spanning trees of GB are in 1 : 1

correspondence with spanning trees of GB. Thus


det(LB) = t(GB). (4.4)

Combining (4.4), (4.2) and (4.1) with (∆) yields

r(i, j) =det(L(i, j))

det(L(i))=t(GA)t(GB)

t(GA)t(GB)= 1,

Whenever the edge (i, j) is a bridge.

Lemma 4.1 supplies a one directional test of whether a graph contains a bridge. Namely,

the absence of unity among the entries of R ensures that there is no bridge. The example

Figure 4.4 below shows that the converse does not hold.

Figure 4.4: The resistance distance between vertex i and vertex j is r(i, j) = 1,however, i and j are not adjacent.

The next function that was considered as a possible indicator of Hamiltonicity was the

gap between the second largest and the largest eigenvalue (λ1−λ2). Although from the

beginning of this chapter, the eigenvalues cannot solely determine Hamiltonicity, they

may still be explored as contributing explanatory variables for the logistic model. It

is known that in a special sense this gap (λ1 − λ2) describes the degree of randomness

and connectivity of the graph [Brouwer and Haemers, 2011, p.67]. For connected cubic

graphs the largest eigenvalue is necessarily equal to three, so studying the second largest

eigenvalue λ2 is equivalent to studying (λ1 − λ2) and this is what will be done in this

study. The second largest eigenvalue has previously been used to relate the so called


toughness3of a graph to Hamiltonicity in Chvatal [1973] and the properties of that

relationship are still being studied. We will illustrate one of the ways the second largest

eigenvalue describes a graph. Consider a d-regular graph G, let X and Y be two subsets

of the vertices of G such that X and Y are disjoint. Let r denote the number of edges

connecting a vertex in X to a vertex in Y . Then it was shown in [Brouwer and Haemers,

2011] that for every possible disjoint X and Y the following holds:

r ≥ (d− λ2)|X||Y |N

,

where |X| is the number of vertices in the subset X. Therefore a small λ2 describes a

cubic graph that has high average ‘connectivity’ throughout the graph and a large λ2

describes a cubic graph that may have ‘bottle-neck’ like features.

Next, we considered a function that is related to the multifilar structure discussed in

Chapter 1 that is called the Estrada index (EE(G)). It was originally introduced as

a measure of connectivity in complex networks and is related to the number of closed

walks of all lengths in the graph [Estrada, 2011]. It is defined as:

EE(G) =n∑i:=1

eλi .

Continuing with the idea of using information about closed walks, the next indicator

function attempts to develop a notion of distance between cubic graphs. A measure

borrowed from information theory called the Kullback-Leibler divergence is tradition-

ally used to measure a difference between discrete probability distributions. Given two

different discrete probability distributions P and Q, the Kullback-Leibler divergence

measures the error when P is used as an approximation for Q and is defined as:

D(P ||Q) =∑i

P (i)ln[P (i)

Q(i)

]

A modification to the formula allows the measurement of ‘divergence’ between graphs.

The matrix exponential of an adjacency matrix A(G) is a matrix whose entries have

3Toughness is another measure of the connectivity of a graph.


contributions from walks of all lengths in the graph, as can be seen in the following

formula,

eA(G) =

∞∑k=0

1

k!A(G)k,

because [A(G)k]i,j is the number of all possible walks of length k that start at vertex i

and end at vertex j. For the purpose of modelling with this modified Kullback-Leibler

divergence a benchmark graph is needed, one which all other graphs can be measured

against and one that exists for all orders of cubic graphs. The answer is found in the

Mobius ladder graph (see Figure 4.4) as it exists for all orders of cubic graphs and it is

also among the most symmetrical of cubic graphs. We now define two versions of the

modified Kullback-Leibler divergence as follows. Let G be a graph of order N and M

be the Mobius ladder graph of order N . Then let E = eA(G) and F = eA(M) and E

and F be the column normalised matrices of E and F respectively. Then the modified

Kullback-Leibler divergence with respect to the Mobius ladder graph is defined as:

D(G||M) =1

N3

N∑k=1

N∑j=1

N∑i=1

[E]ijln[E]ij[M ]ik

And the modified Kullback-Leibler divergence with respect to the graph itself is defined

as:

D(G||G) =1

N3

N∑k=1

N∑j=1

N∑i=1

[E]ijln[E]ij[E]ik

By using the Kullback-Leibler divergence for the matrix exponentials we attempt to

capture a notion of probabilistic distance between the Mobius ladder graph and the

graph in question. The results section will analyse some of the behaviours of this new

index. One reason why the above formula was appealing for use as an explanatory

variable is that this function is only defined for connected graphs. This can be seen by

taking a disconnected graph G, if a vertex i is in one component of G and a vertex j

is in another, then [eA(G)]i,j = 0 and is also zero in the normalized version. Then we

would have ln(0) for some entries in the modified Kullback-Liebler divergence of G and


therefore it is undefined. Another reason the formula was appealing for use is the fact

that it remains constant under isomorphisms of the graph, this fact is left unproven.

Figure 4.5: The Mobius ladder graph of order 10 which is used as a benchmark graphin the modified Kullback-Leibler divergence.

4.2 Discrete valued descriptors of graphs

All of the above functions are investigated for use as explanatory variables in the results

chapter. Graphs can also be described with several functions that take on discrete values

(existence of a Hamilton cycle is one of these). Below is a brief introduction to some

discrete valued functions that were also investigated.

The diameter of a graph is the longest-shortest path between any two vertices. That is,

given a list of the shortest paths (δ(i, j)) between all pairs of vertices in the graph, the

diameter is the maximum of these.

diam(G) = maxi,j∈V (G)

δ(i, j)

The girth of a graph is the length of the shortest simple cycle. Recalling that cycles

of length two are not simple cycles, for cubic graphs the girth is necessarily three or

greater.

A graph is planar if it can be arranged so that its vertices lie on a 2 dimensional plane

and its edges do not cross-over each other (intersect).

The number of triangles present in the graph dictates in which segment the graph belongs

to in the multifilar structure discussed in Chaper 1.


The number of spanning trees in a graph was introduced earlier and it is also considered

for use as an explanatory variable.

Lastly, whether the adjacency matrix possesses a zero eigenvalue (and hence detA(G) =

0) proved to be an interesting descriptor. A zero eigenvalue means that a row or column is

a linear combination of some other rows or columns. This fact means that certain graph

structures force the existence of a zero eigenvalue. For example Figure 4.5 shows two

small subgraphs whose presence in a graph G will force G to possess a zero eigenvalue.

The S labels represent a connection to anywhere in the graph (as long as it remains a

cubic graph). To see that a zero eigenvalue is present, in the left subgraph the vertices

i and j share the same neighbourhoods, so immediately their rows (and columns) are

identical in the adjacency matrix, hence a zero eigenvalue. In the right subgraph, a quick

calculation reveals that, if we let a denote the row in the adjacency matrix for the vertex

a, then a−b−c+d = 0 and hence there is also a zero eigenvalue. So presence of at least

one zero eigenvalue splits all cubic graphs into two groups, one group contains graphs

which possess one or more of such structures (hence det(A(G)) = 0) and the other group

contains graphs which possess none of these structures (hence det(A(G)) 6= 0).

Figure 4.6: Two of the smallest cubic subgraphs that force the adjacency matrix ofthe overall graph to possess a zero eigenvalue. The S label represents a connection toany other part of the graph, including connecting to another S (as long as it remains a

cubic graph).

To conclude this chapter we note that these functions are all interesting in their own

right and some are the subjects of on-going research. Their use as explanatory variables

in this investigation is aimed at a more broad analysis of how they behave when they

are applied to cubic graphs.

Chapter 5

Results

5.1 Local stability with respect to the order of the graph

The goal of the investigation is to construct a model from a random sample of 2 and

3-connected cubic graphs which can predict if a given cubic graph of the same order

is non-bridge non-Hamiltonian. If the explanatory variables are appropriate, the same

model can also be used for orders of graphs that are nearby the order if the graphs

in the original calibration sample. We have some empirical evidence that this, indeed,

the case. To see why this can be so, we need to consider how the non-bridge non-

Hamiltonian graphs behave as the order of the graphs changes. Figure 5.1 demonstrates

the behaviour of the non-bridge non-Hamiltonian graphs for two of the variables intro-

duced in Chapter 4. We observe that, with respect to these variables, as the order of

the graphs increase, the number of graphs with similar characteristics also increases,

that is, there is a clustering effect. In the Figure 5.1 below, observe that the density of

non-bridge non-Hamiltonian graphs is much greater in the cluster located in bottom left

of the plots and the cluster continues to increase in density as the order of the graphs

is increased. When the variables in consideration produce a single main cluster we call

this cluster of non-bridge non-Hamiltonian graphs, including the Hamiltonian graphs

scattered within, a ‘region’ where most of the non-bridge non-Hamiltonian graphs reside

with respect to these variables. If these regions are well enough defined for a smaller

order, then throughout this investigation we have observed that there is a carry-over

effect and this region is also present in the same location for larger orders. The variables

24

Chapter 5. Results 25

that produce a single main region are the ones that prove to be the best explanatory

variables in the next section.

Figure 5.1: Kurtosis of resistances versus variance of resistances for orders 16, 18 and20. Note that the red dots are shown at an increased scale for clarity.

5.2 Results

Earlier chapters have described the construction of a logistic model and the functions of

graphs that may be of interest to use as explanatory variables. This results chapter will

proceed by presenting three cases where a logistic model is constructed and then used

to predict non-bridge non-Hamiltonian graphs for larger orders. As with any statistical


model on more than one predictor variable, the amount of correlation present is assessed

and given in Appendix B. Under this approach a larger correlation value is acceptable

as long as it is not too close to one. The first case is constructed using three predictor

variables namely the second largest eigenvalue, the variance and the kurtosis of entries

in the resistance matrix. All three are certainly related in some way to the structure of

a graph as discussed in Chapter 4. The logistic model will attempt to ascertain to what

extent the three variables together can identify non-bridge non-Hamiltonian graphs.

A random sample of cubic graphs of order 24 is generated using the pairing algorithm

introduced in Chapter 3. The details of this sample are displayed in Table 5.1. From

this calibration sample, a logistic model is constructed and the estimates shown in Table

5.2 are provided by the maximum likelihood estimation technique discussed in Chapter

3, that is, they form the vector β in equation (3.1).

Table 5.1: Information about the population and random sample of cubic, 2 and3-connected graphs of order 24.

Population size 118,118,367

Number of nBnH in population 177,832

Sample size 1,793,560

True number of nBnH graphs in sample 1,246

Table 5.2: Coefficients from the logistic model on the calibration sample of cubic, 2and 3-connected graphs of order 24.

Parameter Variable associated to parameter Estimation of parameter

β0 Constant term -63.763

β1 Variance of resistances -9.956

β2 Kurtosis of resistances -1.770

β3 Second largest eigenvalue 23.258

To both assess the model fit and then use the modelled parameters for predicting non-

bridge non-Hamiltonian graphs for larger orders, first a threshold probability is needed.

This is a probability value that is assigned so that after the model is fit to the sample

data, 90% of the non-bridge non-Hamiltonian graphs in this callibration sample are


correctly assigned an ‘nBnH’ label. Then using this threshold, in order to observe the

model’s fit, we calculate how many of the Hamiltonian graphs are correctly assigned an

‘H’ label . The results for this are displayed in Table 5.3.

Table 5.3: Threshold probability and error values produced for the calibration sampleof cubic, 2 and 3-connected graphs of order 24.

Threshold probability: 0.002345

Correctly predicted Incorrectly predicted

nBnH graphs 89.97% 10.03%

H graphs 84.60% 15.40%

The threshold probability is then fixed and used with the model parameters to predict

for larger graphs, which in this case, are orders 28 and 30. The accuracy of the model for

the larger orders of graphs is then calculated by assessing the error values produced when

applying the threshold to the modelled probabilities. The results for orders 28 and 30 are

displayed in Table 5.4 and 5.5 respectively. It is expected that the accuracy diminishes,

the further away the order gets from order of the graphs in the original calibration

sample. This is true for the proportion of Hamiltonian graphs correctly assigned an ‘H’

label, however, it is observed that the proportion of non-bridge non-Hamiltonian graphs

correctly assigned an ‘nBnH’ label remains approximately constant across the different

order graphs.

Table 5.4: Results when model is used to predict on a random sample of cubic, 2 and3-connected graphs of order 28.

Sample size (order 28) = 598,550

True number of nBnH graphs in sample = 239

True number of H graphs in sample = 598,311

Threshold probability = 0.002345



H graphs 81.14% 19.86%


Table 5.5: Results when model is used to predict on a random sample of cubic, 2 and3-connected graphs of order 30.

Sample size (order 30) = 349,337






H graphs 70.48% 29.52%

To continue this further, we consider a second model constructed again on a sample of

order 24 cubic graphs and using the same explanatory variables, except we discard the

kurtosis of resistances variable. This is because the kurtosis of resistances is observed

to scale with the order of the graphs being considered, which the model cannot account

for. We aim to remove the error introduced by the scaling of this variable and hope that

this increases the accuracy of the predictions. We also narrow the population down to

2 and 3-connected cubic graphs which possess zero triangles (no simple cycles of length

3). This sub-division allows us to restrict the model to graphs that already have one

important structural characteristic in common. It may seem like this is too much of a

simplification in our search for the non-bridge non-Hamiltonian cubic graphs. However,

HCP remains NP-complete even when considering only triangle-free cubic graphs, this

fact is proved in the Appendix A.

A random sample of now triangle-free, 2 and 3-connected, cubic graphs of order 24 is

generated. The details of this calibration sample are displayed in Table 5.6. This time

the number of cubic graphs in the desired population is unknown without generating and

evaluating all of the 118,118,367 order 24 cubic graphs. This means that the estimated

parameters, which are shown in Table 5.7, do not take advantage of the prior correction

that was discussed in Chapter 3. A threshold probability is assigned so that 90% of the

triangle-free non-bridge non-Hamiltonian graphs in the calibration sample are correctly

assigned an ‘nBnH’ label. Again using this threshold, in order to assess the model’s

fit, we can calculate how many of the triangle-free Hamiltonian graphs are correctly

assigned an ‘H’ label. The results for this are displayed in Table 5.7.


Table 5.6: Information about the random sample of cubic, triangle-free, 2 and 3-connected graphs of order 24.

Population size Unknown

Number of nBnH in population Unknown

Sample size 323,838

True number of nbnH graphs in sample 81

Table 5.7: Coefficients from the logistic model on the calibration sample of cubic,triangle-free, 2 and 3-connected graphs of order 24.



β1 Second largest eigenvalue 43.38

β2 Variance of resistances -24.94

Table 5.8: Threshold probability and error values produced for the calibration sampleof cubic, triangle-free, 2 and 3-connected graphs of order 24.




H graphs 94.94% 5.06%

Similarly to the first model, the threshold probability value is then fixed and used

with the modelled parameters to predict Hamiltonicity for graphs of larger order, which

again, are orders 28 and 30. The details of these are displayed in Tables 5.9 and 5.10

respectively. It is observed that the proportion of Hamiltonian graphs that are correctly

assigned an ‘H’ label is significantly higher than in the first case (that is, the second

model produces a lower Type 2 error).


Table 5.9: Results when model is used to predict on a random sample of cubic,triangle-free, 2 and 3-connected graphs of order 28.

Number of graphs in sample = 89,972


True number of H graphs in sample= 89,953



nbnH graphs 89.47% 10.53%

H graphs 89.20% 10.80%

Table 5.10: Results when model is used to predict on a random sample of cubic,triangle-free, 2 and 3-connected graphs of order 30.

Number of graphs in sample = 159,141






H graphs 85.43% 14.57%

These first two cases show that, at least for orders of graphs nearby those of the original

sample, most non-bridge non-Hamiltonian graphs can be identified provided that the

Type 2 error is not of great concern. We can summarise the result of these first two

cases with the statement:

A logistic model can be constructed from cubic 2 and 3-connected graphs of order N

using the second largest eigenvalue, the variance, and the kurtosis of the entries in

the resistance matrix as explanatory variables. The model can identify the region where

most non-bridge non-Hamiltonian graphs reside for orders close to N . The proportion of

Hamiltonian graphs that are correctly identified increases when only triangle-free, 2 and

3-connected graphs are considered and the kurtosis of resistances variable is discarded.

The third case considers two explanatory variables that were chosen because of the way

that they distinguish the co-spectral Hamiltonian and non-Hamiltonian pairs which were

discussed in Chapter 4. The explanatory variables are the skewness of the entries in the


resistance matrix (call this function of a graph G, SR(G)) and the modified Kullback-

Leibler divergence (with respect to the graph itself), KL(G). For the 31 co-spectral

pairs (GH and GNH) that exist in cubic graphs of order 22, all but 3 pairs have the

property that SR(GH) < SR(GNH). Similarly all but 5 pairs have the property that

KL(GH) < KL(GNH). So in a similar manner to the first two cases, this case will

investigate these explanatory variables ability to identify non-bridge non-Hamiltonian

cubic graphs in general.

Again we start with a random sample of 2 and 3-connected, cubic graphs of order 24.

The details of this calibration sample are displayed in Table 5.11.

Table 5.11: Information about the population and random sample of cubic, 2 and3-connected graphs of order 24.

Population size 118,118,367

Number of nBnH graphs in population 177,832

Sample size 896,823

True number of nBnH graphs in sample 631

A logistic model is then constructed on the skewness of resistances and the modified

Kullback-Leibler distance and then a threshold probability value is assigned to assess

the model fit. This time the threshold probability, and hence the predicted region, is

relaxed to include 95% (rather than 90%) of the non-bridge non-Hamiltonian graphs in

the sample data. This is because we observe that a very tight cluster of non-bridge non-

Hamiltonian graphs, with respect to the variables, has been generated in the calibration

sample. The results for these are displayed in tables 5.12 and 5.13.

Table 5.12: Coefficients from the logistic model on the sample of cubic, 2 and 3-connected graphs of order 24 using the variables modified Kullback-Leibler divergence

and skewness of resistances.



β1 modified Kullback-Leibler divergence 46.97

β2 Variance of resistances 1.06


Table 5.13: Threshold probability and error values produced on the sample of order24 graphs using the variables: modified Kullback-Leibler divergence and skewness of

resistances.




H graphs 67.66% 32.33%

Next, the threshold probability is fixed and the modelled parameters are used to predict

Hamiltonicity for random samples of 2 and 3-connected graphs of orders 28 and 30. The

results are displayed in tables 5.14 and 5.15 respectively.

Table 5.14: Results when model is used to predict on a random sample of cubic, 2and 3-connected graphs of order 28.

Number of 28 vertex graphs in sample = 798,112






H graphs 78.32% 21.68%

Table 5.15: Results when model is used to predict on a random sample of cubic, 2and 3-connected graphs of order 30.

Number of 30 vertex graphs in sample = 698,664

Actual number of non-bridge non-Hamiltonian graphs = 160

Actual number of Hamiltonian graphs = 698,504



nbnH graphs 90% 10%

H graphs 82.87% 17.13%

The proportion of Hamiltonian graphs that are correctly predicted starts off lower in this

third case. Unlike the other two cases, the proportion of correctly predicted Hamiltonian


graphs increases as the order of the graphs is increased. This observation suggests that

the proportion of Hamiltonian graphs that lie outside of the identified region is becoming

larger as the order of the graphs is increased. The result of this case can be summarised

with the statement:

A logistic model can be constructed from cubic 2 and 3-connected graphs of order N using

the modified Kullback-Leibler divergence and the skewness of the entries in the resistance

matrix as explanatory variables. The model can identify a region where most non-bridge

non-Hamiltonian graphs reside for orders close to N .

The results will be discussed in the next section and we conclude the results chapter

with the note that the order of the graphs in the calibration samples are chosen because

order 24 is the smallest order in which the whole population of cubic graphs cannot be

generated and analysed easily.

5.3 Discussion

The predictions provided by the logistic models in Chapter 5.2 are given when equation

(3.1) is applied to a given graph and the resulting probability lies either above or below

the pre-specified probability threshold. When the model is used for prediction on a

random sample of graphs, the proportion of non-bridge non-Hamiltonian graphs that

are successfully identified is observed to remain close to that of the original sample used

for model collaboration. This is attributed to the original callibration sample of order

24 graphs appropriately representing the distribution of the main cluster of non-bridge

non-Hamiltonian graphs in the population. When the model is used to predict on a

random sample of graphs of a different order to the original, provided that the order is

nearby that of the original, it is observed that the proportion of successfully identified

non-bridge non-Hamiltonian graphs still remains close to that of the original sample.

This is attributed to the carry-over effect of clusters discussed in Chapter 5.1.

The main result from this investigation is the formulation of a technique which is able

to, with respect to certain explanatory variables, identify the region where most of the

non-bridge non-Hamiltonian graphs reside. These variables do not separate the non-

bridge non-Hamiltonian graphs from the Hamiltonian graphs, rather they contain an

area where the likelihood of a graph being non-bridge non-Hamiltonian is greater. The


list of explanatory variables that were discussed in Chapter 4, did not all possess these

high density regions of non-bridge non-Hamiltonian graphs. The variables that were

chosen were observed to produce the best clustering effect when computing them on

small order graphs of under 20 vertices, of which the whole population can be computed

and analysed directly. The clustering that is present in small order graphs gives a

good indication of how the distribution of the non-bridge non-Hamiltonian graphs will

behave for larger orders. It was observed that if no significant clustering of non-bridge

non-Hamiltonian graphs is present in small orders graphs, then the prediction power of

a logistic model will be poor. It is also worth noting that once the variables are chosen,

the logistic model must be constructed using larger graphs, so that the main cluster of

non-bridge non-Hamiltonian graphs is appropriately dense so that it may be statistically

identified.

We have established that the model’s predictive power for graphs of a larger order (orders

28 and 30 in our examples) is due to the carry-over effect that was discussed in Chapter

5.1. Figures 5.2, 5.3 and 5.4 have the modelled explanatory variables as their axes. They

demonstrate the region that the logistic model is able to identify where most of the non-

bridge non-Hamiltonian graphs reside, as well as the carry-over effect when considering

graphs of a larger order. Of course, if more than three variables are used in the model,

the identified region cannot be easily visualised. Out of the models constructed in this

study, it was observed that using two or three variables produced the best predictions

while still keeping the logistic model statistically sound.

Figure 5.2: The generated sample of 2 and 3-connected graphs, of order 24 usedin the first case example. The region identified by the logistic model where most ofthe non-bridge non-Hamiltonian graphs reside is shown on the left and the non-bridgenon-Hamiltonian graphs and Hamiltonian graphs are shown in separate plots for clarity.

When the entire population of cubic 2 and 3-connected graphs of order 24 was considered

for constructing the models, rather than just a random sample, then the distribution of


Figure 5.3: The samples of triangle-free, 2 and 3-connected graphs, of orders 24 and28 used in the second case example. The region identified by the logistic model where

most of the non-bridge non-Hamiltonian graphs reside is highlighted in green.

Figure 5.4: The samples of 2 and 3-connected graphs, of orders 24 and 28 used inthe third case example. The region identified by the logistic model where most of the

non-bridge non-Hamiltonian graphs reside is again highlighted in green.


non-bridge non-Hamiltonian graphs would is quite different from what is seen in Figures

5.2, 5.3 and 5.4. In the whole population, there are very small clusters of non-bridge

non-Hamiltonian graphs spread throughout the majority of the Hamiltonian graphs (an

example of this can be viewed in Figure 5.1). However, we observe that for the variables

we have selected, even for cubic graphs of moderate size, the density of non-bridge non-

Hamiltonian graphs within the main cluster is far greater than those lying outside of the

main cluster. It is unknown if considering the whole population rather than a sample

would greatly affect the region which is identified by the model (for orders that are

non-trivial, say, greater than 22).

In the last case of Chapter 5.2, the percentage of Hamiltonian graphs correctly predicted

increases when predicting on graphs of a larger order. This suggests that, with respect to

these variables, the proportion of Hamiltonian graphs that lie away from the identified

‘nBnH’ region is increasing. It is possible that, with respect to certain variables, as

the order of the graphs becomes large, there is a region where almost all Hamiltonian

graphs lie and a region where almost all non-bridge non-Hamiltonian graphs lie with some

amount of separation between them. This idea, left as an avenue for future research, is

stated as a conjecture:

Conjecture 1: With respect to certain indicator variables, there is an identifiable amount

of separation between the region where most non-bridge non-Hamiltonian cubic graphs

reside and the region where most Hamiltonian cubic graphs reside.

The three cases presented in Chapter 5.2 show that estimating with this technique can be

tractable for graphs of smaller order. As with many graph theory problems, computation

time becomes prohibitive as the order of the graphs increases. The problem arises here

due to the rarity of non-bridge non-Hamiltonian graphs in the population. For graphs of

larger order, any random sampling process that picks from all graphs in the population

with equal probability, will require an extremely large sample size to obtain a reasonable

amount of non-bridge non-Hamiltonian graphs. For example, a random sample of one

million cubic graphs of order 100, may result in zero non-bridge non-Hamiltonian graphs

present in the sample. However, with this said, all of the algorithms used in the results

section run in polynomial time or less, so the tractability of the technique certainly is

not worse than many problems in graph theory.

Chapter 6

Future work and conclusion

6.1 Future work

The exploratory nature of this study has left several questions that still warrant further

investigation.

The application of statistical methods to general problems that are of NP-complete

complexity is an interesting question. Studying these problems using statistics has the

potential to reveal large scale relationships that may be overlooked when studying them

using only algorithmic or theoretical methods. An avenue that is open for future research

is the investigation of relationships between the regions where a cubic graph resides and

the time required for an algorithm to solve HCP for that graph. That is, can we predict

the computational difficulty of HCP for a cubic graph given the region where the graph

resides, with respect to certain variables? A study similar to this has been conducted

primarily on the Boolean satisfiability problem [Leyton-Brown et al., 2014], which is a

different NP-complete problem and certainly the techniques used by the authors can

be transferred to our case of HCP for cubic graphs. Given what has been observed

throughout this investigation, it would not be surprising if certain regions could be

characterized as being easy and others extremely hard for an algorithm to solve HCP.

An observation that was discussed in Chapter 4 is for specific subgraphs whose adjacncy

matrices possess a zero eigenvalue. This property splits cubic graphs into two groups,

one with

det(A(G)) 6= 0

37

Chapter 6. Future work and conclusion 38

and the other with

det(A(G)) = 0.

A question that may be interesting for an investigation is: is there a polynomial com-

plexity algorithm A such that A : G → G′ and G′ is Hamiltonian whenever G was

Hamiltonian, as well as G′ is non-Hamiltonian whenever G was non-Hamiltonian and

det(A(G′)) 6= 0? This would reduce the problem of HCP for cubic graphs to the case

where the adjacency matrix of the graphs in question all have non-zero determinant and

hence are invertible. If there did exist such an algorithm, the fact that every cubic graph

could be reduced to a graph with an invertible adjacency matrix may allow for more

freedom to study HCP using additional methods from linear algebra.

Lastly, an idea that arose during the investigation was whether it is possible to generate

cubic graphs from a specific subpopulation. For example, can we construct a method

to generate all cubic graphs that lie in the zero filar, that is, cubic graphs with no

triangles? This may be important because there currently exists no way to enumerate

the non-bridge non-Hamiltonian cubic graphs and studying the construction of such an

algorithm may lead to a solution for this enumeration problem. A solution would then

also solve a currently open conjecture of Filar et al. [2014] concerning the prevalence of

cubic bridge graphs:

limN→∞

number of bridge graphs of order N

number of non-Ham graphs of order N

?= 1

That is, does the number of cubic bridge graphs grow much faster than the number

of non-bridge non-Hamiltonian graphs as the order of the graphs increases? Note that

the denominator is all non-Hamiltonian cubic graphs of order N, that is, the number of

cubic bridge graphs plus the number of non-bridge non-Hamiltonian graphs.

6.2 Conclusion

By considering a cubic graph as a member of a population of all cubic graphs of the

same order, this investigation has been able to view the Hamilton cycle problem for

cubic graphs in an unorthodox manner. By identifying a graph from the values that it

provides when certain functions are applied to it, we have investigated the characteristic

Chapter 6. Future work and conclusion 39

behaviours displayed by both the Hamiltonian and non-bridge non-Hamiltonian cubic

graphs. After choosing certain functions that display a clustering behaviour with the

non-bridge non-Hamiltonian graphs, we have formulated a technique that uses statistical

methods to identify a region where, with respect to those variables, most of the non-

bridge non-Hamiltonian graphs reside. Although many Hamiltonian graphs remain in

the identified regions, our conclusion based on the evidence provided in the results

chapter is that the probability that a randomly generated cubic graph is non-bridge non-

Hamiltonian and lies outside the identified region, is very low. These regions are also

shown to remain stable for graphs of orders nearby those used in the original callibration

sample. The last case discussed in Chapter 5.2 shows that when the order of graphs is

increased, there is an increase in the proportion of Hamiltonian graphs that reside outside

the identified region. That is, as the order of the graphs is increased, the amount of Type

2 error decreases. This observation led to a question which is left open as Conjecture

1.

Appendix A

HCP for triangle-free cubic

graphs

Lemma 2: The Hamilton cycle problem is NP-complete when considering only cubic

graphs that lie in the zero filar, that is those that possess zero triangles.

Proof. We need to show that we can convert, in polynomial time or less, HCP for

arbitrary cubic graphs into HCP for triangle-free cubic graphs. So let G be an arbitrary

cubic graph. The problem of finding a Hamilton cycle or determining that one does not

exist is known to be an NP-complete problem [16].

Triangles can be found in any graph in O(m3/2) time [8], where m is the number of

edges. Hence they can be found in the cubic graph G in O((3/2n)3/2) = O(n3/2) time.

Any triangles in G can be reduced using the two operations shown in Figure A.1. The

operations can be repeated until either a complete 4 vertex graph, K4 remains, or a

cubic graph with zero triangles remains. Call the resulting graph G.

Both of the operations described preserve Hamiltonicity in cubic graphs (this can be

seen by tracing the possible paths a Hamilton cycle can take through these structures),

so any Hamilton cycle found in G can be converted back to a Hamilton cycle in G. If G

is determined to be non-Hamiltonian, then G is also non-Hamiltonian.

Therefore we have a polynomial time conversion with respect to the order of the cubic

graph, which converts HCP for an arbitrary cubic graph into HCP for a triangle-free

cubic graph. Therefore HCP for triangle-free cubic graphs is also NP-complete.

40

Appendix A. HCP for triangle-free cubic graphs 41

Figure A.1: Two operations that after being performed repeatedly will reduce anycubic graph to either a K4 graph or a triangle-free cubic graph. The S represents a

connection to any part of the graph.

Appendix B

Correlation values

The first case example in Chapter 5.2 uses three explanatory variables on a random

sample of cubic 2 and 3-coonnected graphs. They are the kurtosis of resistances (KR),

the variance of resistances (V R) and the second largest eigenvalue (SLE). The amount

of correlation present in the sample data is given below:

KR V R SLE

KR 1 0.67 0.9

V R 0.67 1 0.78

SLE 0.9 0.78 1

The second case example uses two explanatory variables on a new set of data which

only consisted of triangle-free cubic 2 and 3-connected graphs. They are the variance of

resistances (V R) and the second largest eigenvalue (SLE). The amount of correlation

present in the sample data is given below:

V R SLE

V R 1 0.77

SLE 0.77 1

The third case example uses two explanatory variables on another random sample of

cubic 2 and 3-connected graphs. They are the modified Kullback-Leibler divergence with

42

Appendix B. Correlation values 43

respect to the graph itself (KL) and the skewness of resistances (SR). The amount of

correlation present in the sample data is given below:

KL SR

KL 1 0.74

SR 0.74 1

Bibliography

[1] “Clay Mathematics Institute millenium prize problems”.

http://www.claymath.org/millennium-problems/p-vs-np-problem, 2015.

[2] P. Baniasadi, V. Ejov, J.A. Filar, M. Haythorpe, and S. Rossomakhine. “De-

terministic ”Snakes and Ladders” Heuristic for the Hamiltonian Cycle Problem”.

Mathematical Programming Computation, 6:55–75, 2014.

[3] R.B. Bapat. “Graphs and Matrices”. Hindustan Book Agency, 2014.

[4] R.B. Bapat, I. Gutman, and X. Wenjun. “A simple method for computing resistance

distance”. Zeitschrift fur Naturforschung A., 58:494–498, 2014.

[5] B. Bollobas. “A probabilistic proof of an asymptotic formula for the number of

labelled regular graphs”. European Journal of Combanitorics, pages 311–316, 1980.

[6] B. Bollobas. “Modern Graph Theory”. Springer-Verlag New York, 1998.

[7] A. Brouwer and W. Haemers. “Spectra of Graphs”. Spriger New York, 2011.

[8] N. Chiba and L. Nishizeki. “Arboricity and subgraph listing algorithms”. SIAM

Journal of Computing, 14:210–223, 1985.

[9] V. Chvatal. “Tough graphs and Hamiltonian circuits”. Discr. Math., 5:215–228,

1973.

[10] M.H. DeGroot and M.J. Schervish. Addison-Wesley Publishing Company.

[11] V. Ejov, J.A. Filar, S.K. Lucas, and P. Zograf. “Clustering of spectra and fractals

of regular graphs”. Journal of Mathematical Analysis and Applications, Volume

333:236–246, 2007.

[12] E. Estrada. “The Structure Of Complex Networks”. Oxford University Press, 2011.

44

Bibliography 45

[13] J.A. Filar, A. Gupta, and S.K. Lucas. “Connected co-spectral graphs are not

necesserily both hamiltonian”. Aust. Math. Soc. Gaz., 32:193, 2005.

[14] J.A Filar, M. Haythorpe, and G.T. Nguyen. “A conjecture on the prevalence of

cubic bridge graphs”. Discussiones Mathematicae Graph Theory, 30:175–179, 2014.

[15] M.R. Garey and D.S. Johnson. “Computers and Intractability, a guide to the theory

of NP-completeness”. W. H. Freeman and Company, page 13, 1979.

[16] M.R. Garey, D.S. Johnson, and R.E. Tarjan. “The planar Hamiltonian circuit

problem is NP-complete”. SIAM Journal of Computing, Volume 5:704–714, 1976.

[17] I.M. James. “History of Topology”. Elsevier BV., 1979.

[18] G. King and L. Zeng. “Logistic Regression in Rare Events Data”. Political Analysis,

Volume 9:137–163, 2001.

[19] K. Leyton-Brown, H.H. Hoos, F. Hutter, and L. Xu. “Understanding the empirical

hardness of NP-complete problems”. Communications of the ACM, 57:215–228,

2014.

[20] R. Robinson and N. Wormald. “Almost all regular graphs are hamiltonian”. Oxford

University Press, 2011.

[21] A. Steger and N.C. Wormald. “Generating random regular graphs quickly”. Com-

binatorics, Probability and Computing, Volume 8, 1999.

Documents

Investigating Hamilton cycles using logistic regression